Ollama

Choose local or cloud execution

Ollama provides local model execution and cloud model access. Its official quickstart describes downloads for macOS, Windows and Linux, followed by a choice of a local model or signing in for cloud models.

For local use, you can download a model and start a chat on your computer. The documented local API does not require an API key. Cloud API use requires an account and key, but does not require an Ollama installation. Decide which route you need before assuming that a request stays on your computer.

Model selection also affects hardware requirements. The quickstart explains that larger context windows need more memory and that using system RAM instead of sufficient VRAM can slow responses. Check the chosen model's requirements rather than interpreting download availability as a performance guarantee for your laptop.

Connect a coding agent or desktop app

The quickstart documents launching Claude Code, Codex CLI and OpenCode through Ollama. For OpenAI Codex, it says to install the CLI first, then launch it and select an Ollama model. This is a documented integration, not a claim that an open model is an OpenAI model.

On macOS, the Apps control can connect Claude or ChatGPT Desktop. The ChatGPT Desktop connection applies to Codex mode; regular Chat and voice continue using the usual ChatGPT connection. Check the relevant client setup instead of expecting all conversations in the desktop app to switch models.

A useful first task is to ask the agent to explain a repository's organization. Compare that answer with the files before allowing changes. This is a proposed evaluation, not a hands-on speed or coding-quality assessment.

Build an application with the API

Ollama documents its own chat API and compatibility routes for OpenAI and Anthropic clients. The quickstart says each compatibility route covers a subset of the original API. Check supported features before moving an existing application across.

The documented Responses route is stateless: send conversation history with each request rather than relying on previous-response identifiers. Keep cloud API keys on your application server, outside browser code and source control.

Local execution and hosted processing have different operational costs and data paths. Inspect the selected service's terms and allowance before deploying. The homepage's throughput and coding-score comparisons are publisher claims, not independent measurements made for this entry.